The Lexical Hypothesis & the Big Five Model

ResearchArticle ISSN 2641-4317 International Journal of Psychiatry Research

The Lexical Hypothesis & the Big Five Model Aram Cargill*

*Correspondence: Aram Cargill, Director, Change Challenge and the Adaption Apex Lab, Australia. Director, Change Challenge and the Adaption Apex Lab, Australia. Received: 29 July 2019; Accepted: 25 August 2019

Citation: Aram Cargill. The Lexical Hypothesis & the Big Five Model. Int J Psychiatr Res. 2019; 2(4): 1-4.

ABSTRACT In the last decades, personality has become a major area of research in psychology. At the present time, the Big Five model is the most popular approach to personality, having been used in settings that include research, education, and employment. The HEXACO model, on the other hand, is a more recent approach to personality that includes an additional personality factor, namely Honesty-Humility (H).

The lexical approach on which the Big Five and the HEXACO models are developed does not come without limitations. On the assumption that measuring personality traits with language-based questionnaires is particularly problematic for the Honesty-Humility factor given the stigma associated with concepts indicating dishonesty, we propose an AI-based model for detecting deceptive responses during the assessment of the H factor. Our model is an adaptation of deception analysis reasoning engine (DARE), an AI system that uses high- and low-level facial features to predict deception in real-life trial videos.

Significance Almost a century later, Goldberg [2] developed the lexical The development of AI-based deception detection models may be hypothesis, which assumes, among others, that the differences an important milestone in the advancement of personality testing, between individuals that are socially relevant have been encoded especially in the case of psychological tests where deceptive in language and that cross-cultural factor analysis of dictionary answers are more likely to occur. words that refer to individual differences can be used to find the most important dimensions of personality. Following the lexical The deception reasoning engine we refer to as OceanH2.0 is meant hypothesis, the Big Five Model was developed, which includes to improve the detection of deceptive responses during personality the personality dimensions Extraversion, Emotional Stability/ assessment and, consequently, increase the accuracy of personality Neuroticism, Agreeableness, Conscientiousness, and Openness to tests. Our work is focused on the H factor of the HEXACO model, Experience [3]. an approach to personality based on the Big Five model. The Five-Factor Model has become the most popular approach We believe that our results will inspire future works that 1) to personality, as many studies support the reliability and validity advance the field of AI-based deception detection and 2) find new of the model in assessing personality and making predictions areas where the proposed model can be applied to detect deceptive based on it. For example, a longitudinal study suggested that behaviours with a higher accuracy rate. personality traits and general mental abilities predict several facets of career success over a span of decades [4]. For instance, Keywords high conscientiousness was associated with intrinsic professional HEXACO model, Big Five model, Dark triad. success while extrinsic professional success was associated with low neuroticism, low agreeableness, high extraversion, high Introduction conscientiousness, and high cognitive ability. The origin of the modern study of personality research can be attributed to the development of the lexical approach by Francis Galton [1], who counted the number and types of words used to The HEXACO Model describe a specific type of character. Some studies using the lexical method have revealed the existence Int J Psychiatr Res, 2019 Volume 2 | Issue 4 | 1 of 4 of a sixth personality factor, namely Honesty-Humility (H), while an average of 78.95%. also giving a different interpretation of Emotional Stability and Agreeableness [1]. These developments have led to the emergence Krishnamurthy, Majumder, Poria, and Cambria [8] developed of the HEXACO model of personality. an automatic deception detection model that combines video, audio, and text with Micro-Expression Features. The Micro- The H factor is defined by traits which include greed- Expression Features were extracted from data containing 39 avoidance, sincerity, fairness, modesty, greediness, slyness, and facial micro-expressions provided by Perez-Rosas, Abouelenien, pretentiousness [5]. In this model, Emotionality does not include Mihalcea, and Burzo [6] The proposed model was tested with 121 the hostility elements that are present in the Big Five model, as videos and obtained a deception detection accuracy of 96.14%. hostility is associated with low Agreeableness. The elements The performance of the model was largely influenced by visual associated with sentimentality, which in the Big Five model belong and textual features, which suggests facial display features and to the Agreeableness dimension, are associated with Emotionality unigrams are more important in detecting deception in videos. in the HEXACO model. Deception Analysis Reasoning Engine (DARE) The Honesty-Humility factor, while not universally embraced by Wu, Singh, Davis, and Subrahmanian [9] proposed an automated the personality community, has been shown to explain variance in deception detection system that was used with real-life courtroom several antisocial behaviours, such as those related to psychopathy, trial videos. The system used classifiers trained on low-level narcissism, Machiavellianism, and egoism, as well as prosocial video features to predict human micro-expressions and show that behaviours such as cooperation [1]. predictions of high-level micro-expressions can be used as features to predict deception. The performance of the training system was AI-Based Deception Detection enhanced by fusing the score of classifiers trained on Improved Automatic deception detection has gained substantial popularity Dense Trajectory (IDT) features and high-level micro-expressions in the last few years, as advances in artificial intelligence (AI) has and by using Mel-frequency Cepstral Coefficients (MFCC) from made possible combining modalities such as video, audio, and text the audio domain. to detect deceptive behaviours in real life videos. Developing AI models that can be used to detect deception could be an important DARE obtained an area under the precision-recall curve (AUC) milestone in the advancement of fields such as forensic, which of 0.877 when evaluated on individuals that were not part of the have relied on physiological measurements of deception for training set and outperformed acknowledged models that use decades. The same applications could be used to detect deceptive human annotations of micro-expressions by 5%. When used with behaviours during psychological testing, potentially increasing human annotations of micro-expressions, DARE provided an the accuracy of psychological tests that are more likely to induce AUC of 0.922. deceptive answers. The results obtained by DARE suggest that a vision system based Several models of AI-based deception detection have been on high- and low-level features can provide a higher accuracy of proposed. For instance, Pérez-Rosas, Abouelenien, Mihalcea, and deception prediction compared to humans and complementary Burzo [6] built a multimodal deception detection system designed information from audio and transcripts can further improve the to differentiate between truthful and deceptive statements made by performance. defendants and witnesses during trials. The proposed model was tested on 121 videos and achieved accuracy levels between 60 and Methodology 75% with a model that extracts and fuses features from gesture and Sample linguistic modalities. It is worth mentioning that the authors also We started training the model with 54 volunteers from our own provided a human deception detection study where they compared social network (family members, friends, and colleagues). The the human capacity to detect deception in trial hearings with that AI was later trained with 400 participants recruited from Amazon of the system they developed; the proposed automated system MTurk. For the next phase of the project, we seek to recruit 4000 provided a relative percentage improvement of up to 51%. participants.

Jaiswal, Tabibu, and Bajpai [7] proposed an automatic deception Feature Extraction detection model with real-life visual and verbal data extracted from The input sources are videos where individuals make truthful or 121 videos. Their approach used Open Face with facial action unit deceptive statements. From the video input, the machine extracts recognition to analyse the movement of facial features of witnesses the transcripts, audio, and video (Figure 1). Global Vectors when they are asked questions and Open Smile to analyse acoustic for Word Representation (Glove) [10] is used as the transcript data. Spoken words were analysed with a lexical analysis that features, MFCC [11] is used as the audio features and Improved emphasis pauses and utterance breaks that are sent to a Support Dense trajectory (IDT) [12] as the video features. Vector Machine to test deceit or truth prediction. Utterance-based fusion of visual and lexical analysis was incorporated with a string- Glove is a log-bilinear regression model for unsupervised learning based matching. The method of feature-level of fusion worked at of word representations which has provided good results on word Int J Psychiatr Res, 2019 Volume 2 | Issue 4 | 2 of 4 analogy, similarity, and name entity recognition tasks. Glove participants. In the next phase of the project, the accuracy of the represents words with vectors and uses word co-occurrence model will be tested with 4000 participants. A satisfactory level of statistics for training. The MFCC features model hearing by accuracy will support the usefulness of the model in personality adjusting the frequencies output from the DFT into a Mel scale and assessment and might be an important milestone in the introduction is used for speech recognition. Finally, the IDT model computers of AI in the world of personality research and assessment. local feature correspondences in consecutive frames and uses RANSAC to estimate the camera motion. Following camera Mission Statement motion cancellation, IDT samples feature points densely at several We started this work because we wanted to develop new ways spatial scales and track them through a limited number of frames. through which machines could gain a better understanding of human behaviour at the individual level. As the world is gradually Procedure transitioning from the information age to the age of augmentation, Humility and honesty are more difficult to assess compared to the our digital identities grow in importance and we must understand Big Five personality traits, as responders are less likely to provide what this means in terms of personality development. responses that reflect their actual behaviours. Respondents can provide deceptive responses intentional, due to the cultural stigma A special focus of this project is on what has been referred to as associated with being dishonest, or unintentional, as cognitive “the dark triad”, which refers to personality traits associated with biases may prevent people from recognizing dishonesty to stable maladaptive behaviours, namely those indicating narcissism, themselves. Machiavellianism, and psychopathy. It has been suggested that these behaviours are encouraged and exacerbated by technology, In response to the limitations associated with assessing humility which might be explained by the fact humans can now digitally and honesty, we used a two-tiered system of evaluation and construct their identities and their digital identities can take scaling framework to add the H factor to the Big Five scale. More precedence over the non-digital ones. If digital identities do not specifically, the user was first asked to either confirm or deny a only express conscious desires and choices but also unconscious statement, which allowed the detection of potentially deceptive ones, negative behavioural tendencies left in “the dark” might responses. The user was then asked to either concur or negative become evident in some types of externalized events with negative the same statement. consequences. This assumption encourages new approaches to understanding personality and predicting behaviour. In order to ensure a higher accuracy level in the assessment of the H factor, an observational approach based on physiological The proposed model is not only mean to detect behaviours from parameters was used via artificial intelligence. More specifically, the dark triad but also behaviours indicating traits belonging to the we have used an adaptation of DARE. triad of well-being.

First, the user is asked to either confirm or deny a statement. This allows the detection of potentially deceptive responses. The user is asked to repeat, to either concur or negate the statement.

Figure 1: Deception detection architecture.

We believe that the social stigma associated with behaviours suggesting dishonesty makes an accurate assessment of the H factor on a language-based questionnaire more difficult than with other personality factors. In order to address this potential limitation of the HEXACO model, we have designed an AI-based model of deception detection based on DARE.

At the present time, our model has been trained with over 400 Figure 2: The triad of well-being and the dark triad. Int J Psychiatr Res, 2019 Volume 2 | Issue 4 | 3 of 4 As shown in Figure 2, the triad of well-being consists of empathy, psychology. 1990; 59: 1216-1229. humility, and honesty. It is worth noticing that the triad is 3. Goldberg LR. Language and individual differences: The search conceptualized as containing the traits that are, in relative terms, for universals in personality lexicons. Review of personality opposed to the ones making the dark triad. Our goal is to make and social psychology. 1981; 2: 141-165. the proposed AI-based deception detection model to not only 4. Judge TA, Higgins C A, Thoresen CJ, et al. The big five detect behaviours that lead to the development of digital identities personality traits, general mental ability, and career success where the traits from the dark triad manifest but also to find and across the life span. Personnel psychology. 1999; 52: 621-652. encourage behaviours that enable the development of the traits 5. Ashton MC, Lee K, De Vries RE. The HEXACO Honesty- forming the triad of well-being. Humility, Agreeableness, and Emotionality factors: A review of research and theory. Personality and Social Psychology We do not believe that traditional personality testing can be used Review. 2014; 18: 139-152. to detect the traits belonging to the dark triad, as the assessment 6. Pérez-Rosas V, Abouelenien M, Mihalcea R, et al. Deception of such traits require the use of words that indicate behaviours that Detection using Real- life Trial Data. are stigmatized and this can have a significantly impact on the 7. Jaiswal M, Tabibu S, Bajpai R. The truth and nothing but accuracy of the answers. the truth: Multimodal analysis for deception detection. In 2016 IEEE 16th International Conference on Data Mining For this reason, we use AI in order to identify attitudes and Workshops (ICDMW). 2016; 938-943. emotions based on visual, auditory, and transcript inputs. This 8. Krishnamurthy G, Majumder N, Poria S, et al. A deep learning approach provides an increased accuracy of trait profiling and can approach for multimodal deception detection. arXiv preprint be used to detect behaviours pertaining to the dark triad before arXiv. 2018; 1803.00344. they result in negative consequences. 9. Wu Z, Singh B, Davis LS, et al. Deception detection in videos. In Thirty-Second AAAI Conference on Artificial Intelligence. Conflict of Interests 2018. We will have to mention the paper is published by authors 10. Pennington J, Socher R, Manning C. Glove: Global vectors for associated with the company owning Ocean H2.0 If you will use word representation. In Proceedings of the 2014 conference on your name as the author of the paper, you will have to indicate you empirical methods in natural anguage processing (EMNLP). are CEO at the company producing the AI model. 2014; 1532-1543. 11. Davis S, Mermelstein P. Comparison of parametric References representations for monosyllabic word recognition in 1. De Vries RE, Tybur JM, Pollet TV, et al. Evolution, situational continuously spoken sentences. IEEE transactions on affordances, and the HEXACO model of personality. acoustics, speech, and signal processing. 1980; 28: 357-366. Evolution and human behavior. 2016; 37: 407- 421. 12. Wang H, Oneata D, Verbeek J, et al. A robust and efficient 2. Goldberg LR. An alternative description of personality: the video representation for action recognition. International big-five factor structure. Journal of personality and social Journal of Computer Vision. 2016; 119: 219-238.

Int J Psychiatr Res, 2019 Volume 2 | Issue 4 | 4 of 4